Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Prefix Caching- fix t4 triton error #2517

Merged
merged 5 commits into from
Feb 16, 2024
Merged

Prefix Caching- fix t4 triton error #2517

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
15421a6
fix triton kernel block size for t4 gpu
caoshiyi Jan 20, 2024
2f25efa
format
caoshiyi Jan 20, 2024
30151a7
minor
caoshiyi Jan 20, 2024
47ea3ff
move cap check inside context_attention_fwd
caoshiyi Feb 15, 2024
1824cdb
clean
caoshiyi Feb 15, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
format
  • Loading branch information
@caoshiyi
caoshiyi committed Jan 20, 2024
commit 2f25efa1130197b1f67fef736a6d4c258e84cfae
2 changes: 1 addition & 1 deletion examples/offline_inference_with_prefix.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@
sampling_params = SamplingParams(temperature=0.0)

# Create an LLM.
llm = llm = LLM(model="baichuan-inc/Baichuan2-13B-Chat", tensor_parallel_size=4, enforce_eager=True, dtype="half", trust_remote_code=True)
llm = llm = LLM(model="facebook/opt-125m")

generating_prompts = [prefix + prompt for prompt in prompts]

Expand Down
5 changes: 2 additions & 3 deletions vllm/model_executor/layers/triton_kernel/prefix_prefill.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,6 @@
import torch
import triton
import triton.language as tl
import time

TESLA = 'Tesla' in torch.cuda.get_device_name(0)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

would it be possible to check for compute capability instead? also, we should do this inside context_attention_fwd, as calling CUDA APIs before we set CUDA_VISIBLE_DEVICES will lead to errors.

Copy link
Collaborator

@esmeetu esmeetu Jan 20, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe we can set prefix_block_size as a parameter in CacheConfig and allow user configure in LLM?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this sort of a thing should be ideally derived automatically.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@Yard1 @caoshiyi Does the block size affect the memory utilization or prefix speed?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@esmeetu The block size is mainly dependent on the shared mem size for different GPU architectures. It will affect the prefix-prefill kernel speed a little bit but has nothing to do with the GPU memory utilization.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we set this variable in a function instead of a global variable? Setting it in global variable may lead to issues in distributed setting.


Expand Down Expand Up @@ -621,7 +620,7 @@ def context_attention_fwd(q,
b_ctx_len,
max_input_len,
alibi_slopes=None):

BLOCK = 128 if not TESLA else 64
# shape constraints
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
Expand Down Expand Up @@ -729,4 +728,4 @@ def context_attention_fwd(q,
num_warps=num_warps,
num_stages=1,
)
return
return
Loading